CNN-based modulation classification using image processing techniques for OFDM underwater acoustic communication system

Abstract Automatic modulation classification (AMC) can be used to identify modulation schemes from received signals over underwater acoustic (UWA) communication channels. Depending on the state of the channel, transmission parameters in UWA channels can be adjusted to enhance system efficiency. The UWA channels present numerous challenges, including multiple long paths, the Doppler effect, and ocean ambient noise (OAN), which make the modulation classification process extremely difficult. Orthogonal frequency division multiplexing (OFDM) systems are commonly used in UWA communication due to their ability to combat multipath fading channels and high spectrum utilization. Deep-learning-based AMC has recently achieved rapid development, recently. In this paper, we propose a deep convolutional neural network (CNN) model based on either ResNet50, AlexNet, or basic CNN, which are trained on preprocessed constellation diagrams with sharpening, thresholding, and decimation for modulation classification. Moreover, we use an OFDM system based on the discrete Fourier transform (DFT), cyclic prefix, and equalization over the UWA communication channel. The modulation types used include 4QAM, 8QAM, 16QAM, 32QAM, BPSK, 8PSK, and 16PSK over a signal-to-noise ratio (SNR) ranging from − 5 to 25 dB for classification. Simulation results for the proposed CNN-based model show that the classification accuracy is 99% for the basic CNN with an improvement of 45%, 94% for AlexNet with an improvement of 29%, and 93% for ResNet50 with an improvement of 31%, compared with the benchmark models at the SNR of 10 dB.